Barely a metric from the Football Analytics realm has received attention in a similar way as the Expected Goals-metric (xG). Most TV broadcasters include the cumulative xG values of both teams at half and full-time. Coaches and pundits alike comment on the xG values, attempting to attribute unfavorable results to bad luck rather than to bad skill. The widespread use of xG in TV shows, on websites, and in locker rooms brings the need for reliable and trustworthy xG models. Only if they draw a better picture of how a match unfolded than “raw” goals they will be embraced by the professional football community. Case studies that successfully use the xG metric for real-world problems build trust in the method. For example, Expected Goals form a valid basis for judging coach dismissals (Flepp and Franck 2020) and enable better predictions of future results than goals (Mead, O’Hare, and McMenemy 2023). But besides proving the general validity of the xG construct, it is imperative to create models that produce reliable estimates of scoring probabilities.
This challenge is essentially a mathematical one and since Expected Goals models are almost exclusively built using Supervised Machine Learning algorithms, it seems straightforward to adopt time-tested approaches from Machine Learning to evaluate the models. However, this is exactly the mistake that this post is meant to expose and we will see in a minute why it is not a good idea to view Expected Goals as a Classification Problem.
Short description of classification algorithms and example of perfect separation on the xg problem
short description of classification evaluation metrics
References
Citation
@online{klemp2024,
author = {Klemp, Maximilian},
title = {Expected {Goals} Is Not a {Classification} {Problem!}},
date = {2024-05-18},
url = {https://maxk92.github.io/posts/2024-05-18-Expected-Goals-is-not-a-Classification-Problem/},
langid = {en}
}